Download presentation
Presentation is loading. Please wait.
1
Network calculus ( system theory )
Jean-Yves Le Boudec ICA, EPFL CH-1015 Lausanne
2
Contents 1. background material
2. the greedy shaper viewed as a min-plus system 3. min-plus operators and a theorem 4. the packetized shaper 5. other examples
3
Part 1: Background Material Arrival and Service Curves
Internet integrated and differentiated services use the concepts of arrival curve and service curves
4
Arrival Curves Arrival curve R(t) -R(s) (t-s) Examples:
leaky bucket (u) = ru+b standard arrival curve in the Internet (u) = min (pu+M, ru+b) bits slope r b slope m M time
5
Arrival Curves can be assumed sub-additive
Theorem: can be replaced by a sub-additive function sub-additive: (s+t) (s) + (t) concave => subadditive
6
Service Curve System S offers a service curve to a flow iff for all t there exists some s such that R R*(t) R(s) R* s t
7
The constant rate server has service curve b(t)=ct
buffer s t Proof: take s = beginning of busy period: R*(t) – R*(s) = c (t-s) R*(t) – R(s) = c (t-s)
8
The guaranteed-delay node has service curve dT
seconds T R R* 0 T T (t) Function T t
9
The standard model for an Internet router
rate-latency service curve bits R T seconds
10
Part 2: The Greedy Shaper viewed as a Min-Plus System
fresh traffic R R* shaped traffic s-smooth shaper: forces output to be constrained by greedy shaper storesdata in a buffer only if needed examples: constant bit rate link (s(t)=ct) ATM shaper; fluid leaky bucket controller Pb: find input/output relation
11
A Min-Plus Model for Shapers
fresh traffic R x shaped traffic s-smooth Shaper Equations: (1) x x (2) x R
12
Solving the Min-Plus Model
(1) x x (2) x R Solving the Min-Plus Model Theorem: There is a maximum solution; it is equal to R Proof: (1) find a solution: fixed point x0 = R ; xi = xi-1 and x* = inf {x0, x1, ..., xi, ...} here: = and thus xi = R =x* (2) if x is a solution, then x R thus x R
13
I/O of Greedy Shaper for any shaper, output R
fresh traffic R R* shaped traffic s-smooth for any shaper, output R R is wide-sense increasing, thus R also thus: the greedy shaper output is R*= R
14
A consequence: Greedy Shaper Keeps Arrival Constraints
fresh traffic constrained by a R R* re-shaped traffic The output of the shaper is still constrained by Proof R* = R (R R R*
15
Part 3: Min-Plus Operators and a theorem
G = set of functions R -> R+ that are wide-sense increasing works also if time is discrete: N -> R+ we consider operators : G -> G is isotone if x(t) y(t) => (x)(t) (y)(t) is upper-semi continuous iff infi((xi )) = (infi(xi)) for sequences xi
16
Min-Plus Linear Operators
is min-plus linear if for any constant K, (x + K) = (x) + K (x y) = (x) (y) is upper-semi continuous. Representation Theorem: is min-plus linear <=> there is some H: R x R -> R+ such that (x)(t)=infs[H(t,s)+x(s)] min-plus linear => isotone
17
Other Properties of Operators
is time invariant if for some T y(t) = (x)(t) and x’(t) = x(t+T) => (x’)(t) = y(t+T) is causal if (x)(t) depends only on x(s), 0 s t
18
Two linear operators Convolution by a fixed function: Cs: x -> x s Cs is linear, time invariant, not causal Cso Cs’ = Css’ Idempotent operator hM x-> hM (x) with hM(t) = infst { M(t)- M(s) + x(s) } is idempotent: hMo hM = hM linear, causal, not time invariant
19
The Packetizer Define function PL by PL(x) = L(n) L(n) x < L(n+1) [Chang 99] call PL the operator: PL(R)(t) = PL(R(t)) accumulates bits until entire packets can be delivered PL is idempotent, not linear, but is isotone and upper-semi continuous
20
A Min-Plus Theorem Implicitely contained in Baccelli et al, “Synchronization and Linearity”, Baccelli et al. Theorem: Assume that is isotone and upper-semi-continuous. The problem x(t) b(t) (x)(t) has one maximum solution in G, given by x*(t) = (b)(t) (Definition of closure) (x) = inf {x, (x), (x), (x),...} in other words: x0 = b ; xi = (xi-1) and x* = inf {x0, x1, ..., xi, ...}
21
Part 4: packetized shaper
fresh traffic R R* shaped traffic s-smooth same as previous, but releases only entire packets example : leaky bucket controller Pb: find input/output relation of packetized greedy shaper
22
Model for packetized shapers
fresh traffic R R* shaped traffic s-smooth Define L(i) = l1 + l2 + … + li The output satisfies: (1) R* R* (2) R* R (3) R* is L-packetized
23
Modelling packetized greedy shapers
system equation : R* PL (R*) C(R*) R maximum solution: R* = PL C(R) th 4.3.3: closure ((PId)o(QId))= closure(PQ) thus closure(PL C)=closure(PL o C) after some algebra: R* = inf {R(1) , R(2) , R(3), …} with R(i) = PL o C o … o PL o C (R) i.e. R(0)=R, R(i) = PL(R(i-1) )
24
Numerical Example for R* = PL C(R)
(t) = 25 t/T for t >0, else 0 - smooth<=> at most 25 data unit per time unit R(t)= a burst of 10 packets of size 10 at time 0 R(i) = PL(R(i-1) ) R* = R(4) 1 2 3 4 5 R(1) R(2) R(3)
25
Special Case Theorem (LeBoudec, Sigmetrics 2001) If s = s0 + l with l lmax then PL o C o … o PL o Cs = PL o C oPL and thus R* = PL(R) Applications: if is concave and (0+) lmax then the packetized shaper can be realized as the concatenation : shaper + packetizer leaky bucket controllers based on bucket replenishment are functionally equivalent to leaky bucket based on virtual finish times
26
Part 5: Other Examples Ex3: Variable Capacity Node
fresh traffic m(t) R R* node has a time varying capacity µ(t) Define M(t) =0t m(s) ds. the output satisfies R* R R*(t) -R*(s) M(t) -M(s) for all s t and is “as large as possible”
27
Variable Capacity Node
fresh traffic m(t) R R* R* R R*(t) -R*(s) M(t) -M(s) for all s t R* R hM(R*) thus there is a maximum solution in G, and R*= hM(R) now hR is idempotent thus hM = hM finally: R*(t) = infst { M(t)- M(s) + R(s) }
28
Ex 4: Loss System node with service curve b(t) and buffer X
fresh traffic R(t) buffer X Loss L(t) b R’(t) R*(t) node with service curve b(t) and buffer X when buffer is full incoming data is discarded modelled by a virtual controller (not buffered) fluid model or fixed sized packets Pb: find loss ratio
29
Model for Loss System fresh traffic R(t) buffer X Loss L(t) b R’(t) R*(t) R’(t) satisfies R’ (X + P(R’)) h R(R’) d0 where P is the transformation R’ -> R* assume P isotone and usc (« physical assumptions »); thus R’ = (X + P ) h R(d0) we don’t know P but P Cb theorem: P P’ => P P’ thus R’ (X + Cb ) h R(d0)
30
Representation of Loss
fresh traffic R(t) buffer X Loss L(t) b R’(t) R*(t) we have shown: R’ (X + Cb ) h R(d0) compute the closure, obtain R’, thus the loss process L=R-R’
31
Bound on Loss Ratio Theorem: if R is a-smooth, then L(t)/R(t) 1 - r
with r = min(1, inf t 0 [b(t) +X] /a(t)) best bound with these assumptions proof: define x(t) = r R(t) x satisfies the system equation: x (X + x b) h R(x) (X + P(x)) h R(x) R’ is the maximum solution => x(t) R’(t) for all t
32
Ex 5: Optimal smoothing R(t-D) R’(t) R*(t) Network Smoother ß(t) s Video display B server R(t) Network offers a service curve b to flow R’(t), Smoother delivers a flow R’(t) conforming to an arrival curve s. Video stream is stored in the client buffer B read after a playback delay D. Pb: which smoothing strategy minimizes D and B ?
33
System Equations (1) R’ is s-smooth (2) (R’(t) R(t-D)
Network Smoother ß(t) s Video display B server R(t) (1) R’ is s-smooth (2) (R’(t) R(t-D) R’(t) = 0 for t 0 Define min-plus deconvolution (a Ø b)(t) = sups 0 [a(s+t)-b(s)] x y <=> x Ø y
34
Max-Plus operators replace by , min-plus becomes max-plus
Deconvolution with a fixed function x -> x Ø a is max-plus linear
35
A max-plus model for Example 5
R(t-D) R’(t) R*(t) Network Smoother ß(t) s Video display server R(t) R’ satisfies: (1) R’ R’ Ø s (2) R’ (R Ø )(t-D) a max-plus system, , with minimum solution x* = inf {x0, x1, ..., xi, ...} x0 (t) = (R Ø )(t-D) xi = xi-1 Ø s thus R’ = (R Ø ) Ø s (t-D) = R Ø ( s) (t-D)
36
Example s b(t) R (s b) R(t) minimum value of D 5 * 106 4 3 2 1
Frame # 5 4 3 2 1 * 106 R (s b) R(t) minimum value of D
37
Minimum Playback Delay
D must satisfy : R Ø ( s) (-D) 0 this is equivalent to D h(R, s)
38
Back to our example D=h(R,s b) s b(t) R(t) 5 * 106 4 3 2 1
Frame # 5 4 3 2 1 * 106 s b(t) D=h(R,s b)
39
R(t) D = 435 ms (s b)(t) D = 102 ms (s b)(t) 100 200 300 400 2000
4000 6000 8000 10000 R(t) 10 20 30 40 50 60 70 (s b)(t) D = 435 ms 100 200 300 400 2000 4000 6000 8000 10000 10 20 30 40 50 60 70 (s b)(t) D = 102 ms Ce théorème est le pendant du théorème de capacité
40
Deconvolution is the time inverse of convolution
(4) Invert time again bits R(t) R( ) T (2) S(t) in inverted time bits R() T R(t) bits R(t) R( ) T (1) R (t) in real time bits R() T S(t) S s b R (s b ) (3) Shape with s b s b
41
Back to our example R (s b) R(t)
Frame # 5 4 3 2 1 * 106 R (s b) Infocom 2001: Joint Smoothing and Source Rate Selection (Verscheure, Frossard, Le Boudec) R(t)
42
Conclusion Network Calculus is a set of tools and theories for the deterministic analysis of communication networks Application of min-plus algebra Does not supersede stochastic queueing analysis, but gives new tools for analysis of sample paths Book and slides available online at Le Boudec’s or Thiran’s home page
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.